Manage AVS indexes

AVS offers an extensive set of configuration options for tuning the performance, storage, and recall of your index. This guide offers examples of the most common configurations to consider. For a complete set of configuration options you can either review the types package in our Python documentation or review the help output by running asvec index create --help.

Index Modes

A key configuration setting for your index is the mode you want your index to operate in. There are two specific modes to utilize.

• Distributed - This is the default index type. It keeps your index searchable while you stream in changes to your index. Use this index type if you’re not sure what mode you need.
• Standalone - Standalone indexing builds your entire index in memory. Once your index is finished being built stand alone it will be automatically set to distributed.

Tip

Use a standalone index if you are building an index from scratch, or rebuilding an index for a new embedding model. This builds the index much faster before automatically making the index available for search.

Fixed configurations

The following configuration parameters affect the storage and graph of your index. They cannot be updated once the index is created. You will need to create a new index if you want to change any of these values.

Description

Parameter	Description
Namespace	Namespace for creating the index, which must already exist in your Aerospike cluster. See configuring Aerospike Database for details about setting up a namespace for your index.
Sets	Specifies where your data is stored in Aerospike Database. Uses the null set by default.
Index name	Name of the index. Used primarily for performing searches.
Dimensions	Number of dimensions (length) of the vector. See generating vector embeddings for details about determining the number of dimensions in your vector.
Vector distance metric	Distance calculation used by your index. Options include: SQUARED EUCLIDEAN — Reports squared Euclidean (L2) distance, avoiding the square root. This is still monotonic as the Euclidean distance, but if exact distances are needed, an additional square root of the result is needed. COSINE - Measures the cosine of the angle between two vectors to determine how similar their direction is, without regarding magnitude. DOT PRODUCT - Takes into account angle similarity and vector magnitude. MANHATTAN - Sums the absolute values of the differences between vector components. HAMMING - Measures the number of dimensions where vectors differ.
Vector field	Field name of the vector in the record.
Mode	Indicates if the initial index should be built standalone.

Python

avs_client.index_create(
 namespace="avs-index",
 name="search-space",
 vector_field="img_vector",
 dimensions=768,
 vector_distance_metric=types.VectorDistanceMetric.COSINE,
 sets="index-set",
 mode=types.IndexMode.STANDALONE,
)

Full documentation.

asvec

asvec index create \
 --index-name search-space \
 --namespace avs-index \
 --set index-set \
 --vector-field img_vector \
 --dimension 768 \
 --distance-metric COSINE \
 --index-mode STANDALONE

Full documentation

Caution

The value of vector field is limited to 15 characters. Using a name longer than 15 characters will cause errors.

HNSW parameters

Adjusting HNSW parameters is the primary way to improve the recall of your searches. Keep in mind improving recall comes at the expense of higher latency, and more storage requirements. Because these paramaters affect the graph you can not update them after creating your index.

Description

Parameter	Description	Default
M (max edges)	Number of bidirectional links created per level during construction. Larger values lead to higher recall but slower construction and larger storage requirements.	16
EF	Size of the dynamic list for the nearest neighbors (candidates) during the search phase. Larger values lead to higher recall but slower searches and higher resource utilization.	100
EF construction	Size of the dynamic list for the nearest neighbors (candidates) during index construction. Larger values lead to higher recall but slower index build and higher resource utilization.	100

Python

avs_client.index_create(
 namespace="avs-index",
 name="search-space",
 vector_field="img_vector",
 dimensions=768,
 vector_distance_metric=types.VectorDistanceMetric.COSINE,
 sets="index-set",
 mode=types.Mode.STANDALONE,
 namespace="avs-index",
 name="search-space",
 index_params=types.HnswParams(
  m=32,
  ef_construction=200,
  ef=400,
 ),
)

Full documentation.

asvec

asvec index create \
 --index-name search-space \
 --namespace avs-index \
 --set index-set \
 --vector-field img_vector \
 --dimension 768 \
 --distance-metric COSINE \
 --index-mode STANDALONE
 --hnsw-m 16  \
 --hnsw-ef 100  \
 --hnsw-ef-construction 10000 \

Full documentation

Dynamic Configurations

The following items do not affect index construction or storage, and can be modified as updates after an index is created and can be used to change the performance of your index by adjusting the cache and healer settings.

Caching

The primary way to reduce the latency of your searches is to adjust the cache settings of your index.

Description

Parameter	Description	Default
Index cache max entries	Maximum number of index records held in the cache.	2,000,000
Index cache expiry	Cache expiration time in milliseconds. Set to -1 to never expire.	-1 (no expiry)
Record cache max entries	Maximum number of vector records held in the cache.	0 (off)
Record cache expire	Seconds to keep record values in cache. (-1 for infinity)	0 (off)
Vector integrity check	Check whether cached vector records have been modified before including them in results. Setting this to false will improve latency but could include stale results.	true

Python

avs_client.index_update(
 namespace="avs-index",
 name="search-space",
 hnsw_update_params=types.HnswIndexUpdate(
  index_caching_params=types.HnswCachingParams(
   max_entries=2000000,
   expiry=-1,
   vector_integrity_check=true,
  ),
  record_caching_params=types.HnswCachingParams(
   max_entries=2000000,
   expiry=-1,
  ),
 ),
)

Full documentation.

asvec

asvec index update \
 --index-name search-space \
 --namespace avs-index \
 --hnsw-index-cache-expiry -1
 --hnsw-index-cache-max-entries 2000000
 --hnsw-record-cache-expiry -1
 --hnsw-record-cache-max-entries 2000000
 --hnsw-vector-integrity-check true

Full documentation

Healer

These values control the healing process associated with your index on indexer or mixed nodes. The values tell the cluster how aggressively to manage the healing process for a particular index. By default these values have minimal impact (<20%) on throughput and query performance.

Description

Parameter	Description	Default
Parallelism	Specifies additional threads used by the healer on a single node. Increasing this will increase the amount of CPU spent towards healing.	1
Schedule	A cron expression for running the healer process, set to run every fifteen minutes by default.	Every 15 minutes

Python

avs_client.index_update(
 namespace="avs-index",
 name="search-space",
 hnsw_update_params=types.HnswIndexUpdate(
  healer_params=types.HnswHealerParams(
   parallelism=1,
   schedule="*/15 * * * *",
  ),
 ),
)

Full documentation.

asvec

asvec index update \
 --index-name search-space \
 --namespace avs-index \
 --hnsw-healer-parallelism 1 \
 --hnsw-healer-schedule "*/15 * * * *"

Full documentation

Additional configuration

The following are additional configurations that can be helpful.

Description

Parameter	Description	Default
Labels	Stores information about the index (for example, the model used to create the vector embedding). It is not relevant to search behavior.	N/A

Python

avs_client.index_update(
 namespace="avs-index",
 name="search-space",
 index_labels={"model-used": "CLIP"},
)

Full documentation

asvec

asvec index update \
 --index-name search-space \
 --namespace avs-index \
 --index-labels model=CLIP

Full documentation

Dropping an index

If you no longer need to search across your data, you can drop an index to free up storage. The delete process is handled by the healer based on the original configuration of the index:

Tip

You can back up index data. See backup and restore instructions.

Python

avs_client.index_drop(
 namespace="avs-index",
 name="search-space",
)

asvec

asvec index drop --index-name INDEX_NAME --namespace NAMESPACE

Note

Dropping an index only frees up space for the index records which is handled gradually by the healer. To delete all underlying vector records, you must delete individual records or delete the set that stores your data.

Troubleshooting

We recommend that you use the asvec CLI tool for troubleshooting your index.

Standalone indexing is not finishing

Standalone indexing requires the following two things to succeed:

• A standalone indexer node role available in your cluster, and

• Sufficient resources on that node to fit the index in memory.

  If either of these conditions are not met the index status will not become READY.

1. Verify that there is a STANDALONE INDEXER node.

   asvec node ls
   ╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
   │ Nodes   │
   ├───┬────────────────┬────────────────────────┬─────────────────────┬─────────────────────┬──────────┬───────────────────────────────────────────┤
   │   │  NODE │ ROLES   │ ENDPOINT   │ CLUSTER ID │ VERSION  │ VISIBLE NODES  │
   ├───┼────────────────┼────────────────────────┼─────────────────────┼─────────────────────┼──────────┼───────────────────────────────────────────┤
   │ 1 │ 37627287442312 │ [INDEX_QUERY] │ 34.56.201.26:5000   │ 1685524865657087021 │ 1.1.0 │ {  │
   │   │ │   │   │   │ │  37912119677832: [34.123.26.109:5000]  │
   │   │ │   │   │   │ │  37956902785928: [34.133.135.181:5000] │
   │   │ │   │   │   │ │  39450770215816: [35.225.89.37:5000]   │
   │   │ │   │   │   │ │ }  │
   ├───┼────────────────┼────────────────────────┼─────────────────────┤   ├──────────┼───────────────────────────────────────────┤
   │ 2 │ 37912119677832 │ [INDEX_QUERY] │ 34.123.26.109:5000  │   │ 1.1.0 │ {  │
   │   │ │   │   │   │ │  37627287442312: [34.56.201.26:5000]   │
   │   │ │   │   │   │ │  37956902785928: [34.133.135.181:5000] │
   │   │ │   │   │   │ │  39450770215816: [35.225.89.37:5000]   │
   │   │ │   │   │   │ │ }  │
   ├───┼────────────────┼────────────────────────┼─────────────────────┤   ├──────────┼───────────────────────────────────────────┤
   │ 3 │ 37956902785928 │ [STANDALONE_INDEXER]   │ 34.133.135.181:5000 │   │ 1.1.0 │ {  │
   │   │ │   │   │   │ │  37627287442312: [34.56.201.26:5000]   │
   │   │ │   │   │   │ │  37912119677832: [34.123.26.109:5000]  │
   │   │ │   │   │   │ │  39450770215816: [35.225.89.37:5000]   │
   │   │ │   │   │   │ │ }  │
   ├───┼────────────────┼────────────────────────┼─────────────────────┤   ├──────────┼───────────────────────────────────────────┤
   │ 4 │ 39450770215816 │ [INDEXER INDEX_UPDATE] │ 35.225.89.37:5000   │   │ 1.1.0 │ {  │
   │   │ │   │   │   │ │  37627287442312: [34.56.201.26:5000]   │
   │   │ │   │   │   │ │  37912119677832: [34.123.26.109:5000]  │
   │   │ │   │   │   │ │  37956902785928: [34.133.135.181:5000] │
   │   │ │   │   │   │ │ }  │
   ╰───┴────────────────┴────────────────────────┴─────────────────────┴─────────────────────┴──────────┴───────────────────────────────────────────╯

2. If you have a STANDALONE-INDEXER node, use the following command to confirm the status of your index. Check the STATUS column for NOT-READY.

   asvec index ls
   ╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
   │   Indexes │
   ├───┬────────────────────────┬───────────┬────────────────────┬───────────────┬────────────┬───────────────────┬──────────┬────────────────┬───────────┬────────────┬─────────────┬──────────┤
   │   │ NAME │ NAMESPACE │ SET │ FIELD   │ DIMENSIONS │ DISTANCE METRIC   │ UNMERGED │ VECTOR RECORDS │ SIZE   │ UNMERGED % │ MODE* │ STATUS   │
   ├───┼────────────────────────┼───────────┼────────────────────┼───────────────┼────────────┼───────────────────┼──────────┼────────────────┼───────────┼────────────┼─────────────┼──────────┤
   │ 1 │ sift-128-euclidean_Idx │ avs-data  │ sift-128-euclidean │ HDF_embedding │  128 │ SQUARED_EUCLIDEAN │  0 │  0 │   0 │   0  │ STANDALONE  │NOT-READY │
   ╰───┴────────────────────────┴───────────┴────────────────────┴───────────────┴────────────┴───────────────────┴──────────┴────────────────┴───────────┴────────────┴─────────────┴──────────╯

3. When the status says READY, run the following command to switch your index mode to DISTRIBUTED to preserve your index updates.

   asvec index update \
   --index-name <INDEX-NAME> \
   --namespace <NAMESPACE> \
   --index-mode DISTRIBUTED

4. After updating to distributed mode, monitor the status of your index by checking to see how much remains unmerged.

   asvec index ls --seeds 35.225.89.37
   ╭──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
   │   Indexes  │
   ├───┬────────────────────────┬───────────┬────────────────────┬───────────────┬────────────┬───────────────────┬──────────┬────────────────┬───────────┬────────────┬─────────────┬────────┤
   │   │ NAME │ NAMESPACE │ SET │ FIELD   │ DIMENSIONS │ DISTANCE METRIC   │ UNMERGED │ VECTOR RECORDS │ SIZE   │ UNMERGED % │ MODE* │ STATUS │
   ├───┼────────────────────────┼───────────┼────────────────────┼───────────────┼────────────┼───────────────────┼──────────┼────────────────┼───────────┼────────────┼─────────────┼────────┤
   │ 1 │ sift-128-euclidean_Idx │ avs-data  │ sift-128-euclidean │ HDF_embedding │  128 │ SQUARED_EUCLIDEAN │   464334 │   548302 │ 549.43 MB │ 81.56%  │ DISTRIBUTED │  READY │
   ╰───┴────────────────────────┴───────────┴────────────────────┴───────────────┴────────────┴───────────────────┴──────────┴────────────────┴───────────┴────────────┴─────────────┴────────╯

Switching between index modes

As shown in the previous example, switching from STANDALONE mode to DISTRIBUTED mode hands over index construction to the index healer, resulting in an eventually consistent index state. Switching from DISTRIBUTED to STANDALONE forces an entire rebuild of your index.

Caution

Switching from DISTRIBUTED to STANDALONE takes your index offline for searching and can result in a poorer quality index. Aerospike recommends creating a new index and switching over to it instead.